A GAME THEORETICAL APPROACH TO NETWORK SECURITY Abstract - This paper studies the network security problem from the perspective of game theory. The paper specifically studies the eavesdropping network intrusion type in which the goal for the intruder is to obtain as much data as possible. The situation is then modeled as a two-person zero-sum game. The payoff of both sides of the game is quantified and the game situations are obtained by analyzing network security states. In addition to the general definition of the problem as a game theoretical model, a simplified specific scenario is modeled, quantified and analyzed. as possible from an information distribution connection between an information/service provider and a client. The client is assumed to be passive when network security is the problem. Therefore the game consists of two players, service provider or the defender, adversary or the attacker, a two person game. 1 - Introduction 2 - Related Work Improvements in internet and World Wide Web technologies have allowed new and inspiring methods for gathering and distribution of information as well as new facilities in establishing communication for people to utilize. As the ease of use of these services is in a never ending increase mode, our everyday chores became more and more dependent on these technologies. This dependency, as a result, brought the need for security of information in providing these services. Game theory is a very useful tool which helps answer the question “What is the best way to play a game?” in strategy games. A network security problem does not only depend on the tools, strategies, etc. available to the defender but it is also highly affected by the actions taken and/or tools used by the adversary. Therefore the network security problem can easily be modeled by using the game theory. Application of game theory to network security has recently been studied by many researchers and it still is a promising research area. This paper studies network security problem based on game theory model in a way that has not received enough attention. Game theoretical approaches to network security problem are mainly focusing on the network administration level. This paper studies a network security problem which is related to more low levels of the network, namely, this paper will use game theory approach to study the eavesdropping attack in which the goal of the adversary is to gather as much information We will study the possibility of apply game theory to network security problem by trying to identify the game for the general case, and study a smaller case with specific strategies, and choices to analyze effectiveness of game theory in studying network security. 3 - The Game We define the game to be the collection of choices, actions and pay-offs between the two players in this game, attacker and defender. Of course this is a simplification of the general problem. A network can be attacked by more than one attacker at any given time. The scenario of the game is as follows; The defender has established a multimedia connection with a client, and sending a video stream. The client is not a part of the game as the client is not expected to cooperate with, or react to other players. The defenders goal is to keep this connection as safe as possible. So, the goal is to allow attacker to capture minimum amount of data from the stream. Any given part of the stream is equally important. Defender uses active probing to detect intrusion. The attacker is trying to capture as much of the stream as possible by intercepting the stream. We assume that the attacker does not have full information of the network, especially the routes available from source to destination. The attacker chooses routers to attack at random with equal probability. In general the game is defined as G = { Ω,S,U }. (1) Probe Identification Ω = {A,D} is the set of players, A = attacker, D = defender. SA = {S1A, S2A, … , SnA } are the set of strategies available to attacker, and SD is the set of strategies of defender likewise. UA = { U1A, U2A, … , UnA } is the utility of attacker for strategy n. 3.1 Attack and Defense Classification For attacker’s goal, only the type of attacks that allow data capturing is useful. Table I summarizes attackers strategies. Strategy Tool Type of Attack Number of targets at a time Explanation Attacker can chose any type of attack (Infecting routers to forward everything to the attacker, exploit vulnerabilities to confuse routers to forward every packet to every destination, etc.) Attacker may choose to attack on or more targets to attack. As a result of such strategies, attacker’s success probability, time to success, etc. will be affected. The defender has the goal to keep data as safe as possible. Table II summarizes defenders available strategies to minimize the risk. Strategy Tool Type of Defense Time Route Explanation Defender may choose to defend by using active probes, or switching the route periodically or non periodically (route dependent), or use a mix of the first two Time for the defender refers to time between probes or time between switching routes Choosing which route to use for transfer Defender may choose probing in a sequential order or may use other intelligent techniques Defender may chose to act upon intrusion detection or try to identify the attacker It is possible to have more possible strategies for both players, but for simplicity Table I and II will suffice for now. 3.2 Utilities Utility in general can be defined as U = G – C, where G is the gain and C is the cost. But, because the relative values of gain and cost are subjective, we will utilize another function for utility. Definition 1: Utility of defender is the quality of session security. Defenders goal is to minimize the amount of data the attacker captures, which is equivalent to maximizing the safe time for the session, and the defender uses probes to achieve this. Therefore the utility function is defined as follows, UD = [Time in seconds the Systems is safe – Time in Seconds System is unsafe] per probes per second (2) Since the goal for the attacker is the opposite, we can choose between two possible utilities for the attacker both of which will have the same effect on the game. Therefore we will choose attackers utility as follows, UA = - U D (3) This choice makes the game a two-person zerosum game. In addition to that, our game is also a game of imperfect information (as players have no information on each other’s moves), and simultaneous (players do not wait for other player to make their moves). 3.3 Solution of the Game According to the existence principle of Nash Equilibrium Theory, a game has at least one mixed strategy Nash equilibrium if the game is a limited game. Given such an attack defense game, if probability distribution of attacker [ pA = {p1A, … prA} ] and defender [ pD = {p1D, … pkD} ] respectively for their strategies SA = { S1A, … , SrA } and SD = { S1D, … , SkD }, where, pi ‘s are probabilities and ∑ pi = 1 for all I’s for both attacker and defender. Then the mixed strategy Nash equilibrium can be found by solving the following equations. VA (pA,pD) = ∑∑ pmApnDUA(SmASnD) and VD (pA,pD) = ∑∑ pmApnDUD(SmASnD) for m 1 to r and n 1 to k. Then, (pA*,pD*) is a Nash equilibrium if for all pA ; VA (pA*,pD*) >= VA (pA,pD*) and for all pD ; VD (pA*,pD*) >= VD (pA*,pD). 4 – Sample Scenario because the router that creates ICMP-TE will add its address to the packet. Therefore, allowing the defender to detect that the packets are out of the route. Of course, for this to be possible, the defender must enforce a specific route at any given time. We are also assuming for this scenario that the defender using a very simple search. Probe routers in the route starting from first to the last and repeat this periodically. 4.2 Assumptions and Definitions We have the following assumptions for the scenario; Defender is using probe detection with periodic probing. Defender is probing sequentially. Defender switches routes upon detection and does not try to identify. Attacker is only attacking one node at a time. Our goal in this paper is to show that game theory can be used to solve security problems in a computer network even at the network layer. In order to study a game theoretical solution to the problem, we will introduce a much simpler version of the game. In the rest of this paper, we will mainly study the strategy of choosing the time intervals between probes for the defender and also the effect of probability of attack for the attacker. Here, the probability of attack should not be considered as a strategy choice for the attacker, but we are rather investigating the effect of assumed attack probability for the defender. Such probability is nothing but an assumption for the defender as this is a game of imperfect information. Ri = Route i, 4.1 Probes Tja = Time to attack node j (with success or fail) We propose a very simple active probing technique for the defender to detect intrusions in this scenario. The probe used by the defender in this game is a special packet that resembles a regular data packet; in this case it is a packet that looks like a video packet. As the attacker attacks routers (making the router forward everything to the attacker or forward everything to every destination), it basically haves the router duplicate everything. In order to detect this duplication, the probe is designed to die before reaching the destination. This is done by giving specific values to TTL of the packet. The goal for using TTL is to create an ICMP Time Exceeded message by the router where the packet is destroyed. Therefore, if the packet is duplicated before purge, the defender will receive more than one copy of the ICMP-TE, thus detecting the intrusion. This probe will work just as fine if the attacker is having the router forward everything to it without duplication Aj = Average time it takes to successfully break into node j Nj = Node j, Mi = Time between probes for route i, P = Probability of attack Tid = Average time to detect an attack on route i Ts = Average time for successful attack 4.3 Calculating the Utility In order to calculate the Utility function as described in section 3.2, we need to make the following calculations. Let P the probability that a single malicious packet is successful in breaking into the router, then; P{Breaking into a node in j th attempt} = (1-P)j-1 P Number of attacks required to break into node: A = ∑ j ( 1-P ) j-1 P = 1/P Therefore, Aj = Tja / P If there are N nodes in total, and r nodes in a route and attacker is choosing any node to attack with equal probability. because it is the only tool for the defender to detect intrusion. Of course this is true for the scenario chosen and because of simplifications to the general problem. To find expected number of nodes until route is found, let X be the number of nodes attacked until route is successfully compromised. To find E[X], let Bi = 1, if i th node out of route is selected before any node in the network, 0, otherwise Then, X-1 = ∑ Bi (sum from 1 to N-r) E[Bi] = 1 / (r+1) Therefore, Figure – 1 E[X-1] = ∑ 1 / (r+1) = (N-r) / (r+1) Figure 1 depicts the relationship between interarrival time of probes and utility of the defender with everything else is fixed. Best utility for the defender is 260 and is achieved at 69 ms between probes E[X] = E[X-1] + 1 = (N-r) / (r+1) + 1 = (N+1) / (r+1) Then, Ts = E[X] * Tja / P = ( (N+1) * Tja) / ( (r+1) * P) (4) Then, for detection, there are r nodes in the route, the probability of a node being compromised given that one of them is compromised is 1/r (i.e. attacker is choosing nodes with equal probability). Neglecting differences in distances between nodes, if detecting an attack on first node is t, then second is 2t, third is 3t, … In this simulation, the size of the network is 50 nodes with an average of 10 intermediate nodes per available route is assumed. Another important observation is worth mentioning in this experient is, as the the network size increases, the size of the routes should increase at a decreasing rate if the same utility is to be achieved. Then, Tid = ∑ kt * 1/r = r * (r+1) * t / (2r) = (r+1)*t / 2 (5) Where t is Mi + (transmission and propagation time). By using (4) and (5) the utility function of the defender is; UD = [Ts – Tid] * Mi (6) 5 Analysis In this section we will analyze the utility function. First of all, the most important factor in our scenario is the time between probes, Mi, Figure – 2 Figure 2 shows a network of size 100 with average route size of 20. This shows that if the average route size to network size ratio stays same as network size increases then the max utility expected for the defender is reduced. Figure 3 – Figure 3 is showing the relationship between assumed value of P, which is the probability of the success of an attack, and Utility value and Mi. Again, the value for P here is the value assumed by the defender in use for strategy determination. It also shows how mistakes in this assumptions will affect the expected utility of the defender. 6 Conclusions In this paper we tried to show that Game Theory can be applied to the network security problem. Even in our very simple scenario, game theory proves to be a useful tool in the decision making process for network security. However, finding the solution to the complete model as given in section 3.3 may not always be possible. The Nash equilibrium is proven to exist for a finite problem. Therefore, in order to guarantee existence of a solution, problem needs to be translated into a finite problem. In our specific case for example, time between probes is not a finite set of available options. But, using some tools, it can easily be turned into a finite set which is a very good representative of the actual set. For example, equipment capabilities can dictate the minimum time between probes and one way to find a maximum value for the same could be achieved by setting a limitation on allowable percent of time the network is in unsafe state.